51³Ô¹Ï

Topic

Load Optimization for Connected Smart Green Buildings Using Machine Learning

Department of Electrical and Computer Engineering

Date & location

• Monday, June 23, 2025

• 9:00 A.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. T. Aaron Gullliver, Department of Electrical and Computer Engineering, 51³Ô¹Ï (Co-Supervisor)

• Dr. Ilamparithi Thirumarai Chelvan, Department of Electrical and Computer Engineering, UVic (Co-Supervisor)

• Dr. Hossen Teimoorinia, Department of Physics and Astronomy, UVic (Outside Member) 

External Examiner

• Dr. Maryam Saeedifard, School of Electrical and Computer Engineering, Georgia Institute of Technology 

Chair of Oral Examination

• Dr. Jon Noel, Department of Mathematics and Statistics, UVic

   

Abstract

Energy efficiency plays a crucial role in mitigating Greenhouse Gas (GHG) emissions, particularly in the building sector, where residential buildings are among the largest energy consumers. Despite the potential of buildings to generate Renewable Energy (RE), increasing energy demand poses both environmental and economic challenges. This dissertation presents an advanced Machine Learning (ML)-based framework for optimizing energy consumption in Connected Smart Green Townhouses (CSGTs) in Burnaby, British Columbia (BC) in Canada, focusing on efficiency, cost-effectiveness, emission reduction, and occupant comfort. 

A comprehensive study on adaptive, occupant-aware, and ML-based energy optimization frameworks presented for Smart Green Townhouses (SGTs) and CSGTs. The work is organized into five integrated research and review studies that collectively address the prediction, optimization, and real-time management of energy consumption, with a focus on sustainability, occupant comfort, and system intelligence. 

Various ML models evaluate for building energy Demand Prediction (DP), revealing that many conventional models lack adequate performance in terms of accuracy and efficiency. A critical review of classification and regression approaches was performed to assess their suitability for modern energy forecasting tasks. 

CSGT complex model operating in grid-connected mode presented. The model incorporates sustainable building materials, smart sensors, Photovoltaic (PV) systems, and energy efficient components. A hybrid Long Short-Term Memory-Convolutional Neural Network (LSTM-CNN) model develope and test on real utility datasets. The results show that this approach consistently outperforms traditional ML models such as Linear Regression (LR), CNN, LSTM, Random Forest (RF), and Gradient Boosting (GB). Metrics such as Mean Absolute Percentage Error (MAPE) below 5%, and coefficient of determination R2 above 0.85 validated the model’s accuracy for different bedroom configurations. 

The dissertation focuses on island mode operation of CSGTs. A robust ML-based optimization framework was proposed for energy and load management under disconnection from the grid. The integration of Electric Vehicles (EVs) with Vehicle-to-Grid (V2G) functionality enhanced system resilience. The LSTM-CNN model again shows superior predictive accuracy with a MAPE of 4.43% and Root Mean Square Error (RMSE) of 3.49 kWh for the 4-bedroom unit. The study confirms that occupant-aware optimization significantly improves operational performance under isolated conditions.  

An adaptive control framework to enable automatic transitions between grid-connected and island modes developed. By incorporating occupancy, weather, and electricity price data, the system dynamically optimized load consumption using LSTM-CNN and Multi Objective Particle Swarm Optimization (MOPSO). Efficiency gains of 3–5% in grid mode and 10–12% in island mode were observed, alongside 4–6% reductions in carbon emissions, demonstrating the value of real-time adaptive management. 

An occupant-centric load optimization system leveraging real-time Internet of Things (IoTs) data proposed. This human-centric approach significantly improve comfort and operational efficiency. Energy loads were reduced by 7–13%, peak loads by 11%, and carbon emissions by 15–24%. Cost savings of 13–21% were achieved, and occupant satisfaction improved with a 19% increase in thermal comfort and 14% better lighting adequacy. 

Collectively, the findings of this dissertation present and highlight the effectiveness of advanced, applicable, and scalable ML-driven energy optimization in SGTs. The proposed frameworks offer scalable, adaptable, and occupant-centric solutions for energy-efficient, cost effective, and environmentally sustainable residential buildings. Future research directions include integrating advanced renewable energy storage management, real-time grid interaction, federated learning models, and edge AI deployment, further enhancing the adaptability and efficiency of smart energy and load management in CSGTs. By integrating advanced ML models, real-time sensor data, and adaptive control techniques, this work provides impactful solutions to address key economic, environmental, and social challenges in sustainable urban housing.